1

The Role of Closed-Loop Hand Controlin Handshaking Interactions

Francesco Vigni1, Espen Knoop2, Domenico Prattichizzo1,3 and Monica Malvezzi1,3

Abstract—In this paper we investigate the role of haptic feed-back in human-robot handshaking by comparing different forcecontrollers. The basic hypothesis is that in human handshakingforce control there is a balance between an intrinsic (open-loop)and extrinsic (closed-loop) contribution. We use an underactuatedanthropomorphic robotic hand, the Pisa/IIT hand, instrumentedwith a set of pressure sensors estimating the grip force appliedby humans. In a first set of experiments we ask subjects to mimica given force profile applied by the robot hand, to understandhow human perceive and are able to reproduce a handshakingforce. Using the obtained results, we implement three differenthandshaking controllers in which we varied the intrinsic andextrinsic contributions and in a second set of experiments weask participants to evaluate them in a user study. We show thata sensorimotor delay mimicking the reaction time of the CentralNervous System (CNS) is beneficial for making interactionsmore human-like. Moreover, we demonstrate that humans exploitclosed-loop control for handshaking. By varying the controllerwe show that we can change the perceived handshake quality,and also influence personality traits attributed to the robot.

Index Terms—Physical Human-Robot Interaction; NaturalMachine Motion; Modeling, Control, and Learning for SoftRobots

I. INTRODUCTION

THE handshake is an important social interaction, commonas a greeting in many parts of the world and in both

business and social contexts [1]. Handshakes contribute to firstimpressions of a person; a drab handshake can have a negativeimpact on the perception of a person’s character. However,there is little work in the literature studying human-humanhandshaking, and as such it is not yet possible to describe whatconstitutes a ‘good’ or a ‘bad’ handshake, or even describe ahuman-human handshake, in a quantitative manner.

In robotics, there are many examples of handshakes used asiconic example interactions for different robotic systems andanthropomorphic hands, often involving a photo opportunity.However, to date there has been relatively little work studying

Manuscript received: September 10, 2018; Revised November 30, 2018;Accepted December 28, 2018.

This paper was recommended for publication by Editor Allison M. Okamuraupon evaluation of the Associate Editor and Reviewers’ comments. This workhas been supported by the European Commission Horizon 2020 FrameworkProgramme, through the Soma project (grant H2020-ICT-645599).

1FV, DP and MM are with Department of Informa-tion Engineering and Mathematics, University of Siena,Siena, Italy. francesco.vigni@student.unisi.it,prattichizzo@ing.unisi.it, malvezzi@dii.unisi.it

2EK is with Disney Research, Stampfenbachstrasse 48, 8006 Zurich,Switzerland. espen.knoop@disneyresearch.com

3DP and MM are with Department of Advanced Robotics, Istituto Italianodi Tecnologia, via Morego 30, 16163 Genova, Italy.

Digital Object Identifier (DOI): see top of this page.

FR

FH

Fig. 1. Human-robot handshaking. While handshaking, human applies on therobot palm a force FH that is measured by a set of pressure sensors, whilerobot applies a force FR to human palm. In this paper we analyze somepossible controllers relating FH and FR.

the handshaking interaction itself. The handshake is alsointeresting in a Human-Robot Interaction (HRI) context [2].In typical HRI tasks, leader and follower roles are clearlydefined: master action and intention is measured and elabo-rated to generate reference inputs for the slave controller. In ahandshake, this identification of roles is not evident a priori,it is an inherently bidirectional action in which both sidesactively contribute to the task by applying an active and areactive action at the same time.

For the case of a human-robot handshake (Fig. 1), the robotwill receive a force from the human which we will denoteFH , and also exert a force FR on the human. The relationshipbetween FH and FR would appear to form a central elementin determining the quality of a handshake. In this work, weare examining this relationship more closely by implementingand comparing different possible controllers relating FH andFR on a robot hand.

The central question which we seek to answer is the roleof feedback in handshaking behavior. We compare open-loopand closed-loop handshake controllers, in order to determinethe importance of haptic feedback and closed-loop control inhandshaking. We use a soft underactuated anthropomorphicrobot hand, the Pisa/IIT SoftHand, instrumented with pressuresensors in order to measure the grasping force exerted ontoit, and we estimate the grasping force exerted by the robotfrom its pose. We then implement three controllers combiningopen-loop and closed-loop to varying degrees. For closed-loopcontrollers, we show that participants perceive interactionsas more human-like if a sensorimotor delay is added tothe system. We objectively demonstrate that humans exhibitclosed-loop handshaking control. In a user study, we evaluatethe subjective perceived qualities of the different controllers

2

and show that the choice of controller can influence the qualityof the handshake as well as perceived personality traits of therobot.

II. RELATED WORK

The area of HRI has received much attention, for examplewith regards to safety [3, 4] and looking at cooperative manip-ulation tasks [5] or handovers [6]. In this work, we focus onhandshaking interactions, which have received limited studyto date.

A. Understanding human handshakes

One of the key elements in realizing a human-robot hand-shake is the measure of the interaction force, in particular,the force that the human applies on the robot palm, indicatedwith FH in Fig. 1. Contact forces in human interactionshave been measured using sensorized gloves, both for objectgrasping [7] and for handshaking interactions [8, 9]. In [10]the authors study human–human handshaking interactions andmeasure contact forces along with IMU hand motion data.In [11] contact area and contact pressure are measured inhuman-human and human-robot handshaking tests, comparingthe results obtained with two underactuated soft hands. Thehandshake grasp would appear to match well to the firstpostural synergy of the human hand, which the Pisa/IITSoftHand has been designed to follow in an underactuatedmanner [12].

Furthermore, handshaking is a complex task from the cog-nitive point of view and poses interesting questions fromthe neuroscientific point of view. For instance, Vanello et al.[13] investigate the neural correlates of human–human andhuman–robot handshake using functional Magnetic ResonanceImaging.

B. Robots for hand interactions

There has been much work in the development of anthro-pomorphic robot hands, such as the Gifu hand [14], the DLRHand 2 [15] or the Shadow Dexterous Hand (Shadow RobotCompany). While human-robot handshakes may often beperformed with various anthropomorphic hands, to showcasethem, there has been limited work where handshaking wasexplicitly studied.

There are some examples of setups for tele-handshaking inthe literature [16]. Pedemonte et al. [17] present the designand realization of a haptic interface performing a robotichandshake, the device is aimed at developing a communicationsystem that allows two people to shake hands while being indifferent locations. Their system for human–robot handshakinginteractions, includes a robot arm controller, a custom handand a hand controller. The design of the system is informed byhuman performance, and the complete system is evaluated in auser study. In follow-up work [16], the same authors considerthe communication system composed of two interfaces andpropose a control algorithm that allows bilateral interactionbetween the two users. Another device for the realizationof realistic human-robot handshake is presented in [18], in

particular a standard model of the human-palm compliance isdeveloped, based on human hand anatomy and an empiricalstudy.

The goal of these systems is to appear as a transparent hapticlink between the two participants, so that the dynamics oftheir interaction is similar to in a direct physical handshake.This is different to our goal, which is to realize a roboticautonomous setup able to emulate the human dynamics inhandshaking tests. Tsalamlal et al. [19] study human-robothandshaking, investigating the effect on perceived affectiveproperties as the arm stiffness, grasping force and robotfacial expressions are changed. Also, Ammi et al. [20] studythe emotional expressiveness of robots combining visual andhaptic interaction (realized through human-robot handshake),and verified that in the identification of some emotional cues,namely dominance and degree of control of a situation, hapticinteraction is more accurate.

Not concerning handshaking, but still mimicking humanhand interactions with a robotic system, Fitter and Kuchen-becker [21] study hand interactions in hand clapping gamesbetween humans and robots.

C. Human consensus dynamics

There are many examples in robotics of modelling humanmotion as dynamical systems, frequently with the goal of thenreproducing similar behavior with a robot.

Humans have been shown to exhibit synchronization, with-out a clear leader or follower, in joint action tasks [22].

Wang et al. [23] propose a haptic virtual reality systemwhich allows human to make physical handshakes with avirtual partner. Two approaches are proposed: in the firstone robot controller employs an embedded curve and dis-regards human interaction, in the second one an interactivecontrol is implemented; they verified that the second oneis perceived more human-like. Karniel et al. [24] proposea Turing-like handshake test to compare a human-humanhandshake, realized through a haptic interface, with differentvirtual handshake models. Both [23] and [24] focus on armtrajectory and disregard handshake force.

D. Human grip strength control

For grasping and manipulation tasks, there are a substantialnumber of studies looking at how the grip force is modu-lated [25, 26, 27], these works show that cutaneous feedbackis used to avoid slip. This principle has also been applied torobotic grasping: Ajoudani et al. [28] propose a system formodulating the grasp strength in a reflexive manner to avoidobject slippage. This is different to a handshaking interaction,and it is therefore not clear to what extent these dynamics willalso be applicable for a handshake.

III. EXPERIMENTAL SETUP

In the experiments presented in this paper, we need tocontrol the force the robot squeezes the human with, FR, andmust be able to measure the force the human squeezes therobot with, FH , as shown in Fig. 1.

3

We use a Pisa/IIT SoftHand for our experiments [12], ananthropomorphic robotic hand based on soft-robotics technol-ogy, exploiting the principles of synergies in an underactuateddesign that is safe for physical human-robot interaction, andadaptable to grasp different objects without any change in thecontrol action.

The Pisa/IIT SoftHand is underactuated, and has a singleactuated degree of freedom q corresponding to the referenceposition of the hand. More specifically, in this paper weindicate with q the main variable that we use to control thehand. It is a variable that ranges from 0 (hand fully opened)to 19000 (hand fully closed). The hand is underactuated andcompliant, so, when it touches an object or a surface, it adaptsto their shapes, and its consequent configuration qa differsfrom the reference one, q. We define q0 as the position wherethe hand is making contact with an object, but applying zeroforce. Once the hand reaches qa = q0, if the object can bemodelled as a rigid body, the actual hand configuration cannotchange. If the object is deformable, increasing the referenceposition results in a relationship between the difference q−q0and the force that the hand is applying to the object, FR, i.e.

FR =

{fR(q − q0) q − q0 ≥ 0

0 q − q0 < 0(1)

where fR is a function mapping the difference q − q0 to theforce FR, that depends on robot hand and object stiffness. Theprocedure that we implemented to identify fR function fromexperimental data and the obtained results will be describedin Sec. IV-A.

The specific value of q0 is dependent on the object that isbeing grasped. This can be obtained in a manual calibrationexperiment, but we have also implemented an automatedcalibration procedure. In the automated procedure, which canbe seen in the supporting video, the robot hand closes slowlywhile the current is monitored. When the robot hand makescontact with the human palm, a steep increase in the currentrelative to the free-closing value is seen. This rise is detected,and the hand position is used to estimate q0.

To evaluate the interaction with the human and measureFH , we attach 3 Force-Sensitive Resistors (FSRs) to the robothand palm (Fig. 2 a)). The FSR sensors have a low profile,so they can be attached to the hand without requiring designchanges. We use the histogram from [11] as a guide for whereto place the sensors, we indicate with Ffsr,i, with i = 1 · · · 3,the measure of the generic sensor. Sensors 1 − 3 in Fig. 2a), are used as triggers to identify the contact with the humanhand, and 1 and 2 are used for estimating FR as they werefound to be robust towards small variations in the grasp.

Although we do not sensorize the entire contact area, we canassume that for similar grips we can estimate FH from the sumof the FSR measurements. To identify the relation between FH

and FSR measurements, we attach the FSR sensors indicatedwith 1 and 2 to a sensorized palm, as shown in Fig. 2 b). Thesensorized palm is a simple 3D-printed object whose shapeand dimensions similar to a human hand palm, composedof two shells connected by a load cell [11]. Six calibrationexperiments were performed, with three different subjects. In

a)

b)

Fig. 2. Experimental setup. a) The Pisa/IIT SoftHand used in this paper,with 2 FSR sensors attached to the side of the palm to measure the forcewhich the human applies to the robot (FH ). b) A sketch of the setup for FSRcalibration, in which sensors are connected to a palm sensorized with a loadcell.

Fig. 3. Plot of Ffsr against FH (measured with load cell), along with best-fitcubic polynomial.

each test, the subject was asked to repeatedly grasp and releasethe sensorized palm, and FSRs and load cell values wererecorded. We then fitted a cubic polynomial to the data, asshown in Fig. 3. This allows us to estimate FH . Although thereis some error in the fit, we observed that for a given handshakegrasp between a participant and the robot the estimate of FH ismonotonic and with relatively low variation—the main sourceof variation is variations in the human grasp configuration.

IV. OPEN-LOOP HUMAN HANDSHAKING DYNAMICS

We now wish to understand the force dynamics of a human-human handshake. In a handshake between participants A andB, participant A squeezes participant B with a force FAB andis squeezed by participant B with a force FBA.

Before contact is made in the handshake, FAB = FBA = 0and other sensory modalities such as vision are relied on.

4

40 42 44 46 48 50 52 54 56 58 600

1020304050

time (s)

F H (N

)

0 20 40 60 80 100 120 140

10000

15000

time (s)

q

40 42 44 46 48 50 52 54 56 58 60

5000

10000

15000

time (s)

q

qqaq0

trial 1trial 2trial 3trial 4trial 5

Fig. 4. Open loop test. Top: reference q profile applied in all the tests. Middle:q, qa, q0 profiles in a part of the experiment. Bottom: estimated FH valuesfor 5 different trials of the same subject, in the same experiment part.

During this phase, we can assume that each participant alsoidentifies a nominal handshake strength to apply based onintrinsic factors such as prior expectation (Fint).

Once contact has been made, the haptic modality becomesdominating (as visual cues of grasping force are minimal). Inthe handshaking phase, each person squeezes with a force andreceives a force. Once cutaneous sensory feedback is available,i.e. after the reaction time of the CNS, we hypothesize thatthe human can be modelled as a dynamical system, and thatthe interaction becomes closed-loop, so that for participant Athe relationship can be expressed as

FAB = f(FBA) (2)

To investigate this relationship, we performed an open-loopexperiment where the robot followed a random fixed trajectory.We asked 8 participants to mimic the grasping force of therobot, as it moved through a random sequence of closuresbetween 6000 and 17000, with each position maintainedfor 3 s. The sequence can be seen in Fig. 4 (top). Eachparticipant repeated the experiment 5 times. Fig. 4 (middle)and Fig. 4 (bottom) show a zoomed-in view of the robotmotion (middle) and the resulting FH for one subject and forthe 5 trials, as measured by the FSR sensors and calibrated asdescribed in Sec. III.

By analyzing the experiments, we noticed a delay in theresponse of 0.2-0.4 s, in almost all the subjects and in mostof force variations. This agrees well with the human responsetime to tactile stimuli [29]. In the following phase, we observethat the human is able to follow robot force variation as quicklyas the variation is applied by the robot. After the transientphase, while the robot is applying a constant force, subjectsshow different behaviors: for example, in some cases we havea force overshoot and a following slower adaptation, and inother cases a steady state case is reached after some oscillation.

The experiments showed that subjects are sensitive to forcevariations; a force adaptation is observed almost each timerobot hand changes FR value. The variation is realized withsimilar reaction time and force rates, however, the reachedsteady state levels exhibit high variability.

The force control employed by a human subject, evenwhen they are asked to mimic a given profile, can be hardlyrepresented as a SISO model as hypothesized in eq. (2), evenif we could identify some analogies in the transient dynamics,the reached equilibrium force contains terms that could not bemodelled in a simple way, such as memory effects, effects oftime, and random variation.

A. Estimating FR

As well as having an estimate of FH , we also require anestimate of the force applied by the robot, FR. It is verydifficult to identify a tactile sensor configuration which wouldestimate FR in a robust way, while not impairing the motionof the hand. Instead, as we are considering handshake graspswith relatively little variation in hand configuration, we assumethat we can estimate FR from the hand configuration, q, andthe point of initial contact, q0.

For this assumption to hold, the setup for the calibrationexperiment should be as similar as possible to a real hand-shaking grasp. For this reason, it is not desirable to use thesensorized palm described earlier.

Instead, we used the data from the open-loop experiment(Fig. 4) as a calibration source. In the experiment, we askedparticipants to match the force of the robot, so if we removethe transients and average across all participants we wouldexpect a quasi-static interaction such that FH ≈ FR. Althoughvariability is observed in the data, we would expect the meanresponse to be robust to noise given a sufficient number oftrials.

For each participant we measured the first contact positionhere noted as q0,j with j = 1 · · · 8. Assuming FR ≈ FH , wethen look for a function

q = f(FH) (3)

In order to obtain a single equation to express the relationin (2), q0,j was used, for which an unique expression can beidentified for qa ≥ q0,j . It is worth to notice that eq. (3) can begenerally expressed for q ≥ q0,j . In order to find this relation,in each experiment we consider FH,j values only for q ≥ q0,j ;where FH,j is the force applied from the j-th participant to therobot hand during the experiment. In this phase we want todefine an average behavior for the force exchanged during ahuman-robot handshake. More formally: in the average modelwe assume that the force FH is hand size independent and isexpressed as:

FH =1

8

8∑j=1

FH,j (4)

Using the Matlab Curve Fitting toolbox, we fitted a cubicpolynomial to the experimental data and obtained a relation-ship between q and FH and therefore approximating q andFR:

q = 0.02 · F 3R − 2.86F 2

R + 157.2FR (5)

5

C1C3bC2b

C3a

C2a

FH

FR

Fig. 5. FR as a function of FH in the proposed controllers. C1: robotfollower, FR = FH . C2: robot open loop, FR and FH are independent;two levels of force are implemented: C2a (lower) and C2b (higher). C3:combined controller FR is dependent both on FH (as in C1) and on therobot’s intrinsically preferred force (as in C2). C3a has a lower intrinsicallypreferred force, and C3b has a higher value.

The whole calibration procedure can be therefore summarizedin two parts: we first use the sensorized palm to expressFH as a function of the FSR measurements, and we thenuse the results from the open-loop experiment to estimate arelationship between FR and q.

V. CONTROLLERS FOR ROBOT HANDSHAKING

We now wish to develop a set of possible controllers forrobot handshaking, exhibiting different possible behaviors.Our goal is to develop a controller for robot handshakingwhich improves the handshake quality. It is reasonable toassume that a robot controller which closely mimics thecontrol rules followed by humans in a handshake will resultin an improved interaction quality.

Considering the haptic interaction part of the handshake, wepropose 3 robot handshake controllers, as described below andillustrated in Fig. 5. We also refer to the accompanying video,where we showcase the behavior of the different controllers.

We note that the robot hand imposes an upper bound on FR,measured to be 50 N. Controller forces are therefore saturatedto this level.

1) Robot follower (C1): A simple controller, and one whichhas been implemented in the literature [17] is to have the robotfollow the human and trying to match FR to FH . This is shownin red in Fig. 5. While this would be expected to produce areasonable force profile over time, it is likely that one wouldperceive the robot as following and responding to the humanmotion. We would expect this to lead to the controller beingperceived as less human-like.

2) Robot open loop (C2): An alternative simple controlleris that the robot is open-loop in the handshaking phase i.e.it sets FR to some value Fint which is independent of FH .Thus, the robot behavior is governed by intrinsic factors ratherthan by extrinsic factors. We consider two versions of this, onewith a lower force Fint = 17.4 N (C2a) and one with a higherforce Fint = 34.2 N (C2b). This is shown in Fig. 5 in blue.

As the controller is not following the human it is more likelyto impose dominance and cause the human to be the follower.Furthermore, we would expect the stronger controller (C2b)to be perceived as more dominant than the weaker one (C2a).

3) Combined controller (C3): It seems reasonable to hy-pothesize that the human controller for handshaking forcecombines both intrinsic and extrinsic information. To imple-ment this, we combine controllers 1 and 2 to get

FR =1

2(Fint + FH)

We sketch this controller in green in Fig. 5. Again, we considera lower (C3a) and higher (C3b) value of F0.

We would expect this controller to be perceived as moreresponsive than C2. If it is the case that humans do combineintrinsic and extrinsic factors, this controller should be per-ceived as more human-like. Similarly to C2, we would expectthe stronger controller (C3b) to be perceived as more dominantthan the weaker one (C3a).

A. Handshake termination

In a human-human handshake, the termination must beinitiated by one participant who reduces the grasping force tozero. For a synchronous termination, the other participant mustdetect this and also release their grasp. Thus, one participantacts as the leader and the other as the follower.

For a robotic handshaking controller, this means that inorder to mimic human behavior the robot should terminate thehandshake if it detects that FH drops below a threshold. Also,the robot should initiate the termination e.g. if the durationexceeds some time limit.

In order to reduce variability in our experiments, we hereonly consider the case of the human leading the termination,and the robot following. If required, it would be straight-forward to implement a time limit on the handshake afterwhich the robot should initiate the termination. As a guideline,Wang [30] report a mean handshake duration of 1.0 s, and amaximum duration of 1.8 s. We note that this could contributeto the robot being perceived as a follower, rather than a leader.

B. Controller delay

It is known that for tactile stimuli, humans have a reactiontime around of 250 ms [29]. For human closed-loop handcontrol, this delay would be expected to be present. Whendesigning a robot controller mimicking the human response,it is therefore reasonable to consider the inclusion of asensorimotor delay in the controller.

To determine whether such a delay would indeed be ben-eficial, and if so how large it should be, we implemented aversion of Controller 1 (robot follower) where a variable delaycould be applied to the sensory signal from the robot tactilesensors. The delay was controlled by a slider in a GUI, witha range from 0 to 300 ms.

We then carried out a study where we asked 5 participantsto shake hands with the robot hand, and to adjust the delaytime to the value which they felt provided the most human-like handshake response. Participants were free to control thedelay time as they wished, and perform as many handshakesas they saw fit, until they found a suitable value.

We found that participants on average preferred a delaytime of 120 ms (standard deviation 90 ms). This shows that

6

Fig. 6. Experimental setup for user study, with the robot hand attached to afixed mount.

the inclusion of a sensory delay, or reaction time, is indeedbeneficial for mimicking human closed-loop actions. The factthat the preferred delay time is somewhat smaller than thehuman response time could be explained by the additionalresponse time introduced by the robot hand.

For the remainder of this work, we therefore implement asensory delay of 120 ms for all the controllers. Note that thedelay is applied to the extrinsic sensory signals (equivalent ofmechanoreceptors), but not to the intrinsic hand dynamics.

VI. SYSTEM EVALUATION

The implemented handshaking controllers have been evalu-ated in a user study, where participants were asked to performhandshakes with the different controllers and rate each one ofthem individually on a set of Likert-scale questions.

The five controllers being tested were introduced in theprevious section and depicted in Fig. 5. All controllers wereimplemented using the sensorimotor delay of 120 ms asdescribed above.

A. Experimental procedure

The robot hand was attached to a rigid mount, as depictedin Fig. 6. For a more realistic test scenario, we did not impairthe vision or hearing of the participants.

15 participants (12 male) were recruited for the study. Theyreceived cinema vouchers in return for their participation. Thestudy was approved by the Disney Research IRB. Participantswere briefed about the study, and asked to sign a writtenconsent form.

For each participant, we first identified q0. This was donemanually, to ensure minimal variation.

Participants were then presented with a randomized se-quence of the 5 handshaking controllers. Each controllerappeared 3 times in the sequence, for a total of 15 trials. Foreach trial, we asked participants to perform a set of handshakes(not a prescribed number) with the robot hand and then answer5 questions as listed in Tab. I. Responses were made on a 7-point Likert scale. The first 3 questions relate to the handshakequality and human likeness, and the last 2 questions relate toperceived personality traits of the robot.

B. Results

1) Handshake statistics: Across all handshakes, we cancompute some statistics. In total, participants performed 1812handshakes (on average 8 per trial), with a mean durationof 2.2 s and with a mean value of FH of 24.8 N. This is

TABLE ILIKERT-SCALE QUESTIONS.

Question Scale (1 to 7)

Q1 Please rate the quality of thehandshake

very poor to very good

Q2 Please rate the human-likenessof the handshake

very robot-like to very human-like

Q3 Please rate the responsivenessof the robot

not responsive at all to veryresponsive

Q4 Who was the leader of thehandshaking interaction

I was the leader to the robotwas the leader

Q5 How would you judge the per-sonality of the robot

shy, hesitant, introvert to con-fident, secure, extrovert

longer than would be expected for a human-human handshake,suggesting that participants might be spending longer time inorder to better understand robot behavior.

2) Do humans follow the robot in C2: For the open-loopcontrollers (C2a and C2b), FR is independent of FH . Todetermine if the human followed the robot in this controller,we computed the mean value of FH across all participants forthe two conditions C2a and C2b. A t-test showed a significantdifference between FH in C2a (M = 19, SD = 10.4) andC2b (M = 27.4, SD = 19.9) with p = 0.0138. This showsthat humans do indeed incorporate closed-loop control forhandshaking, and follow the behavior of the robot.

3) How are different controllers rated: To analyze theresponses from the user study, we first computed for eachparticipant their mean responses for each controller. For eachquestion, we then performed pairwise t-tests with Bonferronicorrection between all pairs of controllers. The results aresummarized in Fig. 7.

For Q1 (handshake quality) we found a significant differ-ence between controllers C2a (M = 3.98, SD = 1.27) andC3a (M = 5.11, SD = 1.09) with p = 0.0012. It can thusbe seen that there is a perceived improvement in handshakequality between the weaker force open-loop controller and theweaker combined controller.

For Q2 (human likeness) we also found a significant dif-ference between controllers C2a (M = 3.62, SD = 1.39)and C3a (M = 4.93, SD = 1.27) with p = 0.0045. Thesame trend as for Q1 is thus seen, with the weaker combinedcontroller being perceived as more human-like than the weakeropen-loop controller. In general, from Fig. 7, it appears thatthere is correlation between Q1 and Q2, as would be expected.

For Q3 (responsiveness) we found significant differencesbetween C2a (M = 3.36, SD = 1.54) and C3a (M = 5.18,SD = 1.32) with p = 0.0022, and between C2a and C3b(M = 4.76, SD = 1.55) with p = 0.0370. The perceivedresponsiveness of the combined controller, both with strongerand weaker force, is therefore significantly greater than thatof the weak open-loop controller. It can be seen that the twoopen-loop controllers are rated as less responsive than the 3closed-loop controllers, as would be expected, however for theremaining pairs this difference is not statistically significant.

For Q4 (leader/follower) we did not find any significanteffects. In general, responses are towards the lower end of thescale meaning that participants felt that they were the leader in

7

1

7

C1 C2a C2b C3a C3b

Q4: who was leader

C1 C2a C2b C3a C3b

1

7

Q1: quality**

1

7

Q2: human-likeness

C1 C2a C2b C3a C3b

**

1

7

C1 C2a C2b C3a C3b

Q3: responsiveness**

*

1

7

C1 C2a C2b C3a C3b

Q5: robot personality***

Fig. 7. Bar charts showing results from user study. Error bars show 95 % confidence intervals. Significant differences between controllers have been indicatedwith ∗ for p < 0.05, ∗∗ for p < 0.01 and ∗ ∗ ∗ for p < 0.001.

the handshake. We note that although there was no significantdifference in leader/follower for C2a and C2b, there was stilla significant difference in FH between the two conditionsmeaning that humans did indeed follow the robot.

For Q5 (robot personality), we found a significant differencebetween C2a (M = 3.09, SD = 1.58) and C2b (M = 4.98,SD = 1.67), with p = 0.00049. For the two open-loopcontrollers, increasing the handshaking force therefore has theeffect of making the robot be perceived as more confident,secure and extrovert while decreasing the force causes it tobe perceived as more shy, hesitant and introvert. To a lesserextent, the same effect can be observed in C3, but in this caseit is not significant.

VII. DISCUSSION

We have shown that humans do employ closed-loop forcecontrol for handshaking, demonstrated with the significantdifference in FH for the open-loop robot controllers withhigh and low force. It is interesting to note that participantsin general still reported that they were the leader of theinteraction, even when the robot did not follow them. This islikely due to the human still being in charge of the initiationand termination of the handshake.

It can be seen that some error is introduced into our systemboth through the estimation of FH and FR. For soft andunderactuated robot hands, grasping force estimation is ingeneral a difficult problem, and there is no readily-availabletactile solution for instrumenting a robot hand for graspingforce estimation. Force estimation from the motor current, aswe do here, inherently suffers from errors due to cable frictionand natural variability in the grasp.

Despite this significant simplification, and the limitations ofour hardware, we have shown that by changing the closed-loophand controller we are able to change perceived qualities ofthe handshake. It would be expected that if other aspects of thehandshake, such as arm dynamics, could also be modulated,then an even stronger response should be elicited.

The results show some directions and preliminary answerson how robot personality is perceived by humans in handshak-ing, for example, with C2b or C3b controllers robot shows amore evident character, while with C1 or C3a it is perceivedas more comprehensive. However, human perception of robotpersonality deserves a more in-depth human-centered analysis,that will be the focus of future works.

Mimicking the delay imposed by the human sensorimotorsystem we were able to improve the perceived quality of thehandshake. It is reasonable to assume that this type of delayed

response could improve the perceived quality of other closed-loop physical human-robot interactions such as collaborativemanipulation tasks. The preferred sensorimotor response timewas found to be 120 ms, while the human tactile reaction timeis around 250 ms. We can attribute the shorter preferred delaytime of the robot to the additional response time added bythe robot hand. It is also plausible that humans incorporatefeed-forward elements to the hand control, predicting whatthe handshaking partner will do, in order to create a shorterapparent reaction time.

In this work, we used the preferred response time for themain user study. However, it would be interesting to furtherinvestigate how the addition of a sensorimotor delay in arobotic system influences the perceived qualities, both forhandshaking interactions and also more generally for othercollaborative tasks.

The robot hand used in this work is anthropomorphicand bears close resemblance to a human hand. It would beinteresting to explore handshaking with simplified robot handsbearing less resemblance to human hands, to see if a similarresponse could be elicited. Looking further ahead, recentwork has demonstrated brain-to-brain and muscle-to-muscleinterfaces for closed-loop human-human interactions [31]. Abetter understanding of human handshaking could pave theway for more realistic handshake-like interactions throughsuch interfaces.

The focus of this work has been on hand control for hand-shaking interactions, however it is clear that there could bemany other factors of robot hand design that will also influencethe perceived qualities, such as hand size, palm compliance,and also hand appearance. Nevertheless, we would expect ourfindings regarding the hand control to generalize to differentrobot hands.

VIII. CONCLUSION

We have objectively shown that humans perform closed-loop hand control during handshaking, with an increase inFR causing an increase in FH . With regards to the subjectiveperceived handshake qualities, our results also suggest thatclosed-loop controllers are preferred. Although we do not havesufficient evidence to conclusively select one of the closed-loop controllers, our results suggest that a combination ofintrinsic and extrinsic control would be preferable.

Moreover, we found that the addition of a sensorimotordelay to the closed-loop robotic system was preferred by par-ticipants for creating a more human-like interaction, analogous

8

to the sensorimotor delay exhibited by humans due to thereaction time of the CNS.

The use of an existing anthropomorphic hand, with min-imal adaptation, means that our controllers could readily beimplemented on robots with anthropomorphic and underac-tuated hands. For a deeper understanding of the closed-loopinteraction dynamics, a specialized non-anthropomorphic testsetup could have produced results with less variability.

We expect that our findings would be relevant for otherclosed-loop haptic interactions such as shared manipulationtasks and human-robot hugging.

ACKNOWLEDGMENT

We thank Nicolas El Maalouly for help with the statisticalanalysis.

REFERENCES

[1] W. F. Chaplin, J. B. Phillips, J. D. Brown, N. R. Clanton, and J. L.Stein, “Handshaking, gender, personality, and first impressions.” Journalof personality and social psychology, vol. 79, no. 1, p. 110, 2000.

[2] T. B. Sheridan, “Human–robot interaction: status and challenges,” Hu-man factors, vol. 58, no. 4, pp. 525–532, 2016.

[3] M. Zinn, B. Roth, O. Khatib, and J. K. Salisbury, “A new actuationapproach for human friendly robot design,” The international journal ofrobotics research, vol. 23, no. 4-5, pp. 379–398, 2004.

[4] C. Gosselin, T. Laliberte, B. Mayer-St-Onge, S. Foucault, A. Lecours,V. Duchaine, N. Paradis, D. Gao, and R. Menassa, “A friendly beastof burden: A human-assistive robot for handling large payloads,” IEEERobotics & Automation Magazine, vol. 20, no. 4, pp. 139–147, 2013.

[5] T. Takubo, H. Arai, Y. Hayashibara, and K. Tanie, “Human-robotcooperative manipulation using a virtual nonholonomic constraint,” TheInternational Journal of Robotics Research, vol. 21, no. 5-6, pp. 541–553, 2002.

[6] W. P. Chan, C. A. Parker, H. M. Van Der Loos, and E. A. Croft, “Ahuman-inspired object handover controller,” The International Journalof Robotics Research, vol. 32, no. 8, pp. 971–983, 2013.

[7] S. Shimizu, M. Shimojo, S. Sato, Y. Seki, A. Takahashi, Y. Inukai, andM. Yoshioka, “The relationship between human grip types and forcedistribution pattern in grasping,” in Advanced Robotics (ICAR). Proc.8th Int. Conf. IEEE, 1997, pp. 299–304.

[8] Z. Wang, J. Hoelldampf, and M. Buss, “Design and performance of ahaptic data acquisition glove,” Proc. 10th Ann. Int. Workshop Presence.,pp. 349–357, 2007.

[9] A. Melnyk, V. P. Borysenko, and P. Henaff, “Analysis of synchronyof a handshake between humans,” in 2014 IEEE/ASME InternationalConference on Advanced Intelligent Mechatronics. IEEE, 2014, pp.1753–1758.

[10] G. Tagne, P. Henaff, and N. Gregori, “Measurement and analysis ofphysical parameters of the handshake between two persons accordingto simple social contexts,” in Intelligent Robots and Systems (IROS),IEEE/RSJ Int. Conf. IEEE, 2016, pp. 674–679.

[11] E. Knoop, M. Bacher, V. Wall, R. Deimel, O. Brock, and P. Beardsley,“Handshakiness: Benchmarking for human-robot hand interactions,” inIntelligent Robots and Systems (IROS), 2017 IEEE/RSJ InternationalConference on. IEEE, 2017, pp. 4982–4989.

[12] M. G. Catalano, G. Grioli, E. Farnioli, A. Serio, C. Piazza, and A. Bicchi,“Adaptive synergies for the design and control of the pisa/iit softhand,”The International Journal of Robotics Research, vol. 33, no. 5, pp. 768–782, 2014.

[13] N. Vanello, D. Bonino, E. Ricciardi, M. Tesconi, E. P. Scilingo,V. Hartwig, A. Tognetti, G. Zupone, F. Cutolo, G. Giovannetti et al.,“Neural correlates of human-robot handshaking,” in RO-MAN, 2010IEEE. IEEE, 2010, pp. 555–561.

[14] T. Mouri, H. Kawasaki, K. Yoshikawa, J. Takai, and S. Ito, “Anthropo-morphic robot hand: Gifu hand iii,” in Proc. Int. Conf. ICCAS, 2002,pp. 1288–1293.

[15] J. Butterfaß, M. Grebenstein, H. Liu, and G. Hirzinger, “Dlr-hand ii:Next generation of a dextrous robot hand,” in Robotics and Automation,2001. Proceedings 2001 ICRA. IEEE International Conference on,vol. 1. IEEE, 2001, pp. 109–114.

[16] N. Pedemonte, T. Laliberte, and C. Gosselin, “A haptic bilateral systemfor the remote human–human handshake,” Journal of Dynamic Systems,Measurement, and Control, vol. 139, no. 4, p. 044503, 2017.

[17] N. Pedemonte, T. Laliberte, and C. Gosselin, “Design, control, andexperimental validation of a handshaking reactive robotic interface,”Journal of Mechanisms and Robotics, vol. 8, no. 1, p. 011020, 2016.

[18] M. Arns, T. Laliberte, and C. Gosselin, “Design, control and exper-imental validation of a haptic robotic hand performing human-robothandshake with human-like agility,” in Intelligent Robots and Systems(IROS), 2017 IEEE/RSJ International Conference on. IEEE, 2017, pp.4626–4633.

[19] M. Y. Tsalamlal, J.-C. Martin, M. Ammi, A. Tapus, and M.-A. Amorim,“Affective handshake with a humanoid robot: How do participantsperceive and combine its facial and haptic expressions?” Proc. 6th ConfAffective Computing and Intelligent Interaction, pp. 334–340, 2015.

[20] M. Ammi, V. Demulier, S. Caillou, Y. Gaffary, Y. Tsalamlal, J.-C. Martin, and A. Tapus, “Haptic human-robot affective interactionin a handshaking social protocol,” in Proceedings of the Tenth An-nual ACM/IEEE International Conference on Human-Robot Interaction.ACM, 2015, pp. 263–270.

[21] N. T. Fitter and K. J. Kuchenbecker, “Using IMU data to demonstratehand-clapping games to a robot,” in Intelligent Robots and Systems(IROS), IEEE/RSJ Int. Conf. IEEE, 2016, pp. 851–856.

[22] C. Zhai, F. Alderisio, P. Słowinski, K. Tsaneva-Atanasova, andM. di Bernardo, “Design of a virtual player for joint improvisation withhumans in the mirror game,” PloS one, vol. 11, no. 4, p. e0154361,2016.

[23] Z. Wang, E. Giannopoulos, M. Slater, and A. Peer, “Handshake: Realistichuman-robot interaction in haptic enhanced virtual reality,” Presence,vol. 20, no. 4, pp. 371–392, 2011.

[24] A. Karniel, I. Nisky, G. Avraham, B.-C. Peles, and S. Levy-Tzedek,“A turing-like handshake test for motor intelligence,” in InternationalConference on Human Haptic Sensing and Touch Enabled ComputerApplications. Springer, 2010, pp. 197–204.

[25] R. S. Johansson and K. J. Cole, “Sensory-motor coordination duringgrasping and manipulative actions,” Current opinion in neurobiology,vol. 2, no. 6, pp. 815–823, 1992.

[26] A. Eliasson, H. Forssberg, K. Ikuta, I. Apel, G. Westling, and R. Johans-son, “Development of human precision grip. v. anticipatory and triggeredgrip actions during sudden loading.” Experimental brain research, vol.106, no. 3, p. 425, 1995.

[27] A. G. Witney, A. Wing, J.-L. Thonnard, and A. M. Smith, “The cuta-neous contribution to adaptive precision grip,” Trends in neurosciences,vol. 27, no. 10, pp. 637–643, 2004.

[28] A. Ajoudani, E. Hocaoglu, A. Altobelli, M. Rossi, E. Battaglia,N. Tsagarakis, and A. Bicchi, “Reflex control of the pisa/iit softhandduring object slippage,” in Robotics and Automation (ICRA), 2016 IEEEInternational Conference on. IEEE, 2016, pp. 1972–1979.

[29] P. Lele, D. Sinclair, and G. Weddell, “The reaction time to touch,” TheJournal of physiology, vol. 123, no. 1, pp. 187–203, 1954.

[30] Z. Wang, “High-fidelity haptics in multimodal human-robot interaction,”Ph.D. dissertation, Technische Universitat Munchen, 2010.

[31] M. E. M. Mashat, G. Li, and D. Zhang, “Human-to-human closed-loop control based on brain-to-brain interface and muscle-to-muscleinterface,” Scientific reports, vol. 7, no. 1, p. 11001, 2017.